Beth Israel Deaconess Medical Center Boston, Massachusetts, United States
Objective: NICU admissions significantly raise neonatal mortality risk and cost up to $168,000. No universal assessments predict NICU admission risk at the point of care. Literature shows early prenatal risk assessment and education/nutrition intervention reduced NICU admissions by 56% and NICU days by 59%. This study aims to develop an ML model predicting NICU admissions using early prenatal variables and demographic/clinical history.
Study Design: ML models were trained on sociodemographic, medical history, and obstetric variables obtainable at the point of care. Models were trained on years 2018-2020 of the CDC Vital Statistics System and were tested on year 2021. NICU patients from years 2014-2017 (with the exception of year 2015 due to an unavailable data file) were included in training to reduce class imbalance.
Exclusion criteria: missing NICU data and non-reporting hospitals. The primary outcome was immediate NICU admission post-delivery. Weight scaling algorithms enhanced training for preterm labor patients and NICU patients from prior years up to 2014 were used to improve class imbalance.
Results: 44 clinical variables obtainable in the 1st trimester across 12,127,124 patients were included in the training data. 2,041,555 (16.83%) experienced NICU admissions.
The model was tested on 3,219,217 patients, of which 351,787 (10.93%) had a NICU admission. The developed model had an area under the receiver operating characteristic curve of 0.76, 77% accuracy, sensitivity of 0.60, and F1 score of 0.77 at a standard 50% threshold.
When Youden’s J statistic was calculated, the optimal threshold was 45.55%, which yielded 0.66 sensitivity.
Conclusion: Machine learning models can identify nearly 2/3rds of all NICU admissions at the point of care using routinely collected factors. This early identification can potentially allow for targeted interventions to prevent over half of all NICU admissions and lead to cost savings for health plans.
